Liquid d1 3B — English Memory Gate

A local decision model designed as an upstream filter for Hindsight and similar conversational memory systems. It returns IGNORE for clearly disposable interaction and CONSIDER for personal information, unresolved references or memory-related requests. CONSIDER means assessment, not automatic storage or permission to execute a command.

Validation

On 57 English test examples from the 296-example curated/synthetic dataset, the selected training adapter scored 56/57 (98.2%) and its F16 llama.cpp export scored 57/57 (100%). One prediction differed between runtimes. These consumed experimental test results are not independent real-conversation validation or a safety guarantee. End-to-end Hindsight integration has not been validated.

License

LFM Open License v1.0 applies; see LICENSE, including the $10 million annual revenue commercial-use threshold. This is not an Apache or MIT license. The base LiquidAI/d1-3B model was modified through English memory-gate LoRA fine-tuning.

This repository: standalone F16 GGUF

The GGUF already includes the fine-tuning. Do not apply an adapter again. Use llama.cpp revision 0878f7c304cc13915c831ae17694534c94940176 or a compatible build with native /v1/systemone support—not chat prompting.

llama-server --model liquid-memory-gate-f16.gguf --host 127.0.0.1 --port 8080 \
  --parallel 1 --ctx-size 512 --batch-size 512 --ubatch-size 512 --n-gpu-layers 99
curl --silent --show-error --fail-with-body http://127.0.0.1:8080/v1/systemone \
  -H 'Content-Type: application/json' --data-binary @request-example.json

Edit only state in the example request. Preserve the trained questions and IGNORE, CONSIDER option order. The label is in answers.memory_gate.choice. There are no Python inference scripts or extra adapter-loading steps.

The full merged safetensors model is in the companion repository.